# V2 contracts manuscript and narration checkpoint — 2026-09-06

This is progress toward [TODOS.md](../../TODOS.md), not library completion. The
delivered catalog remains 38 guides / 380 unique slides / 213 sources. The
unpublished manuscript is outside that catalog, its narration script and
teaching assignments. The immutable source revision remains
`1ceae52c7896a59ac74a4429d410f3218f0b91c6`; incoming `main` changes have not
silently changed the teaching snapshot.

## Authored and source-reviewed scope

The [manuscript](../../drafts/v2-contracts.json) expands four slides to sixteen,
with 21 distinct cited files, explanations and at least three questions per
slide. The sequence follows intent/reference conventions, documentation checks,
target artifacts and module reconciliation, compiler configuration, a worked
determinism contract, cook/DDC routes, static-check limits, generated gRPC
descriptors and a bounded evidence conclusion.

Direct inspection of the pinned implementation qualifies several overview claims
rather than repeating them as established behavior:

- The game and server explicitly disable unity builds; the editor target does
  not contain that override. Its effective setting is not inferred. All three
  classes have separate private determinism and pre-build helper
  implementations.
- The two reserved V2 definitions have no branching consumer in pinned Source
  `.cpp`/`.h` files. Their declaration is not runtime determinism evidence.
- The floating-point checker asserts selected JSON values and source strings.
  Mutation tests show that removing `/fp:strict` fails, while removing unity
  overrides or the contract's `targets`/`automation` fields is not rejected by
  the respective validators. These tests characterize the immutable teaching
  snapshot; they do not require current product code to retain these gaps.
- A fabricated subsection beneath a known parent can pass the bare-reference
  checker. The explicit negative example is illustrative, not an alleged shipped
  broken reference. External HTTP links are skipped by that checker.
- Explicit BuildGraph Cook nodes carry reference-DDC arguments; direct cook
  wrappers do not. The effective standalone wrapper cache remains a question for
  actual execution/environment evidence.
- The proto generator emits sorted C++ metadata descriptors and lookup helpers,
  not serializers or a network transport. Its regex parser has a narrower syntax
  boundary than a general protobuf compiler; byte freshness alone cannot prove
  complete interpretation or live service integration.
- Plugin-directory entries are not equivalent to enabled plugins. In particular,
  the V2Editor resources/widget-contract path is separate from the V2Editor
  Source module and has no plugin descriptor at this pin.

No Unreal runtime, build, cook, simulation or live RPC pass is claimed here.

## Verification and reproduction

Use the [Linux environment](../linux-resume-2026-09-06/README.md), with the
virtual environment first on `PATH` and `CHROME_CHANNEL=chromium`. Run the
commands separately, after repository resource checks, without an owned narrator
running alongside browser verification.

```sh
python tools/presentations/build-eve-oshun.py --draft v2-contracts
python tools/presentations/build-eve-oshun.py --draft v2-contracts --check
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --draft-id v2-contracts --draft
python tools/presentations/build-eve-oshun.py --check --preview
python -m unittest discover -s tools/presentations/tests -p 'test_*.py'
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide portfolio-map
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --controls-only --draft
python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs --smoke --draft
```

The isolated draft builder cannot publish audio/catalog entries, rejects unsafe
identifiers and a mismatched source revision, and retains correct repository
source-link depth. The focused browser mode checks manuscript freshness before
testing and records stable manuscript/HTML hashes. It does not weaken the
whole-bank V1 foundation-source assertion.

Retained evidence:

- [49 passing unit tests](unit-tests.txt), including six pinned-source claim
  tests and two isolated-draft tests. Actual Misaki and MP3 regressions ran.
- [Draft browser report](draft-browser.json) and [log](draft-browser.txt): every
  one of the 16 slides at desktop, compact and mobile sizes; axe, sources,
  transcripts, navigation, dialogs, focus and 16-page PDF sequence/title/footer
  checks. Audio is deliberately absent in this unpublished draft.
- [Published portfolio regression](portfolio-browser.json) and
  [log](portfolio-browser.txt): the full 12-slide suite, including actual audio,
  controls, HTTP failure/race recovery, portable-file operation and PDF checks.
- [Whole-bank controls regression](bank-controls.json) and
  [log](bank-controls.txt): source/transcript alignment, V1 foundation-source
  retention, navigation, dialogs and history. This is not all-slide visual or
  audio acceptance for the 380-slide bank.
- [38-guide center smoke report](center-smoke.json) and [log](center-smoke.txt):
  selection, hierarchy/duration aggregation, search, preparation links, history,
  collapse persistence, performance budgets and mobile drawer/focus behavior.
  Draft smoke mode does not certify unfinished audio or every embedded slide.

Representative direct image inspection covered the desktop target comparison
(slide 8), desktop determinism-contract table (slide 10), mobile DDC flow (slide
12), and PDF page 10 rasterized with Poppler. Text, labels, table rows,
takeaways and footers were readable without clipping in those samples. The
mandatory frontend skill preserved the existing cream/ink/terracotta reading
workspace and chose tables for comparisons and flows for sequences. This sample
review is not the final every-slide/every-PDF visual/editorial audit.

Ruff, scoped ESLint, Prettier, generated-artifact freshness and
`git diff --check` also pass at this checkpoint. Browser reports retain the
actual Chromium version, source pin and artifact hashes.

## Narration progress and remaining work

Nine additional CPU recordings follow the earlier `member-day` recovery:
`member-shell`, `release-enforcement`, `pwa-install-update`, `member-account`,
`member-design-language`, `tara-practice-journey`, `tara-taxonomies`,
`tara-cadence-duration` and `tara-audio-accessibility`.

All **74 current tracks** pass persisted fingerprint, bytes, SHA-256, decoded
duration and encoded-pause checks: **5,673.251 seconds and 887 sentence
segments**. The [audio check](current-audio-check.txt) lists the selected IDs
and manifest hash. **306 remain pending: 287 stale and 19 missing.** Generated
player/catalog readiness and ancestor totals use only current files. No
complete-guide duration was substituted for missing narration.

The initial six-track batch was stopped by its own resource supervisor when
other host work reduced available memory below 2 GiB; the completed recording
was preserved. See the [resource-stop log](narration-resource-stop.txt).
Subsequent fresh one-track CPU processes made progress without changing model,
voice, pronunciation or pause policy. They were deliberately stopped before
browser work; old process IDs must not be reused.

Resume the 306 pending recordings, then run the full inventory and all
applicable player gates. The V2 draft still requires final visual/editorial
review, narration, teaching assignments, prerequisite-linked catalog integration
and published-guide checks. Comprehensive monorepo coverage, all eleven
cross-cutting tracks, Eve integration/ad hoc presentations, Domains and final
visual/listening audits remain part of the full objective.
